The above folder has the code to analyze the Titanic disaster problem using neural network. This dataset performs better with arguably simpler algorithms like Random forests but since this was the first dataset I ever explored, I wanted to start keras in it.
iteration over different NN sizes:
Hidden layer size | train_loss | train_accuracy | dev_loss | dev_accuracy | dropout | Inference |
---|---|---|---|---|---|---|
15 10 10 10 | 0.4592 | 0.8069 | 0.4386 | 0.7982 | no overfitting | |
15 10 | 0.4171 | 0.8204 | 0.4153 | 0.8161 | no overfitting | |
7 5 5 | 0.3919 | 0.8293 | 0.4239 | 0.8341 | no overfitting | |
20 5 | 0.4198 | 0.8278 | 0.4271 | 0.8072 | no overfitting | |
25 5 | 0.3811 | 0.8398 | 0.4522 | 0.8430 | 0.1 0 | 25 best for first layer |
30 5 | 0.3948 | 0.8308 | 0.4024 | 0.8341 | 0.1 0 | |
25 5 5 | 0.3657 | 0.8473 | 0.4108 | 0.8430 | 0.1 0 | even better!!! |